Training Monitor

Configure, launch, and watch training live — with tuning, responsible-AI, and reproducible output.

Live loss / accuracy curves and resource usage
Specialized training tab: text · audio · causal · generative · detection · diffusion · graph · federated · meta · foundation
Autonomous / Neural Architecture Search: define a search space + budget, get the best architecture
Hyperparameter tuning, k-fold cross-validation, external evaluation, uncertainty, and calibration
Responsible-AI: fairness, drift, and carbon-footprint reporting
AI interpretation + publication-quality reports; reproducible code and checkpoints

Annotate images in-browser, then train (no pre-labelled data)

Don’t have a labelled dataset? For the Object detection model, expand "Annotate images in-browser" to open the built-in annotator — upload images, draw, and train in one click. No CSV or folder structure required.

1
Pick a mode
Boxes (object detection), Regions (segmentation), or Labels (image classification).
2
Add images and a class
Click "+ Add images" (any JPG/PNG) and type a class name, then press Enter.
3
Annotate with real tools
A tool rail gives you Draw, Select/Move, Brush (freehand regions), Erase, and Pan, plus Undo/Redo (Ctrl+Z / Ctrl+Shift+Z). For object detection: select a box to drag it, drag its square handles to resize (with a live W×H readout), nudge it with the arrow keys (Shift = 10px), and press 1–9 to set or change its class. Use the thumbnail strip to jump between images (annotated ones are marked) and the mouse wheel to zoom.
4
Train
Click Train — the job queues on the worker fleet and its progress shows inline below. Your images and annotations are saved in this browser and survive a refresh.

The first epoch can take a few minutes

On the current CPU fleet an image or multimodal run spends 1–3 minutes loading the model and data before the first epoch appears, and queued jobs wait their turn. The monitor shows a “Starting up — queued on the training fleet” banner until the first epoch lands; it has not hung. You can leave the page — training continues on the server.

Training runs on a dedicated worker fleet, so the app stays responsive while a model trains.
A run auto-resumes if you navigate away and come back — and the imaging viewer’s Train buttons take you straight to the monitor.
LoRA / parameter-efficient fine-tuning of ViT foundation models is supported.

Prefer pre-labelled data?

Detection accepts a filename,label,x1,y1,x2,y2 CSV; segmentation an images/ + masks/ folder pair; classification class-subfolders. The annotators produce exactly these formats.

AI is privacy-safe

Every AI call sends aggregated statistics only — never raw rows — and runs under anti-hallucination guards.

Architecture Builder

Experiments & Comparison